sglang 0.4.6.post3__py3-none-any.whl → 0.4.6.post5__py3-none-any.whl

sglang/srt/models/mimo_mtp.py

@@ -0,0 +1,220 @@
+# Adapted from https://github.com/vllm-project/vllm/pull/17433/files  and deepseek_nextn.py
+
+from functools import partial
+from typing import Any, Dict, Iterable, Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from sglang.srt.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    split_tensor_along_last_dim,
+    tensor_model_parallel_all_gather,
+)
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import QKVParallelLinear, RowParallelLinear
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.pooler import Pooler, PoolingType
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.models.mimo import MiMoForCausalLM
+from sglang.srt.models.qwen2 import (
+    Qwen2Attention,
+    Qwen2DecoderLayer,
+    Qwen2MLP,
+    Qwen2Model,
+)
+from sglang.srt.utils import add_prefix
+
+
+class MiMoMultiTokenPredictorLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        prefix: str,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.token_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.hidden_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.input_proj = nn.Linear(
+            config.hidden_size * 2, config.hidden_size, bias=False
+        )
+        self.mtp_block = Qwen2DecoderLayer(
+            config=config, quant_config=quant_config, prefix=prefix
+        )
+        self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+
+        if input_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_embeds
+        # masking inputs at position 0, as not needed by MTP
+        hidden_states[positions == 0] = 0
+
+        hidden_states = self.input_proj(
+            torch.cat(
+                (
+                    self.hidden_layernorm(forward_batch.spec_info.hidden_states),
+                    self.token_layernorm(hidden_states),
+                ),
+                dim=-1,
+            )
+        )
+
+        hidden_states, residual = self.mtp_block(
+            positions=positions,
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+            residual=None,
+        )
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layernorm(hidden_states)
+        return hidden_states
+
+
+class MiMoMTP(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        nn.Module.__init__(self)
+        self.config = config
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.quant_config = quant_config
+
+        self.model = MiMoMultiTokenPredictorLayer(
+            config,
+            prefix,
+            quant_config,
+        )
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+        )
+        self.logits_processor = LogitsProcessor(config)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, forward_batch)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head, forward_batch
+        )
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        params_dict = dict(self.named_parameters())
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freq" in name or "projector" in name:
+                continue
+            if "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+            if name.startswith("model.vision_tower") and name not in params_dict:
+                continue
+            name = self.map_model_name_to_mtp_param_name(name)
+
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                if "mtp_block" not in name:
+                    break
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if "mtp_block" not in name and (
+                    "embed_tokens" not in name
+                    and "lm_head" not in name
+                    and "token_layernorm" not in name
+                    and "hidden_layernorm" not in name
+                    and "input_proj" not in name
+                    and "final_layernorm" not in name
+                ):
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+
+    def map_model_name_to_mtp_param_name(self, name: str) -> str:
+        import re
+
+        name_without_prefix = [
+            "token_layernorm",
+            "hidden_layernorm",
+            "input_proj",
+            "final_layernorm",
+        ]
+        pattern = r"model.mtp_layers.(\d+)."
+        group = re.match(pattern, name)
+        if group is not None:
+            for sub_name in name_without_prefix:
+                if sub_name in name:
+                    name = name.replace(group.group(), "model.")
+                    return name
+            name = name.replace(group.group(), "model.mtp_block.")
+        return name
+
+    def get_embed_and_head(self):
+        return self.model.embed_tokens.weight, self.lm_head.weight
+
+    def set_embed_and_head(self, embed, head):
+        del self.model.embed_tokens.weight
+        del self.lm_head.weight
+        self.model.embed_tokens.weight = embed
+        self.lm_head.weight = head
+        torch.cuda.empty_cache()
+        torch.cuda.synchronize()
+
+
+EntryClass = MiMoMTP

sglang/srt/models/minicpmo.py

@@ -1520,12 +1520,15 @@ class MiniCPMO(MiniCPMBaseModel):
         slice_start_id: int = mm_input.slice_start_id
         slice_end_id: int = mm_input.slice_end_id
 
-        media_token_pairs = [
+        data_token_pairs = [
             (im_start_id, im_end_id),
             (slice_start_id, slice_end_id),
             (mm_input.audio_start_id, mm_input.audio_end_id),
         ]
-        pattern = MultiModalityDataPaddingPatternTokenPairs(media_token_pairs)
+        data_start_token_ids = [im_start_id, mm_input.audio_start_id]
+        pattern = MultiModalityDataPaddingPatternTokenPairs(
+            data_token_pairs=data_token_pairs, data_start_token_ids=data_start_token_ids
+        )
 
         return pattern.pad_input_tokens(input_ids, mm_input)
 
@@ -1823,22 +1826,12 @@ class MiniCPMO(MiniCPMBaseModel):
         **kwargs: Any,
     ) -> torch.Tensor:
 
-        mm_input = forward_batch.merge_mm_inputs()
-        placeholder_token_ids = (
-            ([mm_input.im_token_id] + [item.pad_value for item in mm_input.mm_items])
-            if forward_batch.contains_mm_inputs()
-            else []
-        )
         hidden_states = general_mm_embed_routine(
             input_ids=input_ids,
             forward_batch=forward_batch,
             language_model=self.llm,
             image_data_embedding_func=self.get_image_feature,
             audio_data_embedding_func=self.get_audio_feature,
-            placeholder_tokens={
-                Modality.IMAGE: placeholder_token_ids,
-                Modality.AUDIO: placeholder_token_ids,
-            },
             positions=positions,
         )
         return hidden_states

sglang/srt/models/mistral.py

@@ -13,6 +13,12 @@
 # ==============================================================================
 """Inference-only Mistral model."""
 
+from typing import List, Union
+
+import torch
+from transformers.models.mistral3.modeling_mistral3 import Mistral3MultiModalProjector
+
+from sglang.srt.managers.schedule_batch import MultimodalDataItem
 from sglang.srt.models.llama import LlamaForCausalLM
 
 
@@ -20,4 +26,68 @@ class MistralForCausalLM(LlamaForCausalLM):
     pass
 
 
-EntryClass = MistralForCausalLM
+class Mistral3ForConditionalGeneration:
+    MULTIMODAL_PROJECTOR_TYPE = Mistral3MultiModalProjector
+
+    def __init__(self, **kwargs):
+        # lazy load inner class
+        # to bypass circular import
+        from sglang.srt.models.llava import LlavaForConditionalGeneration
+
+        # override config: mistral's projector adds patchmerger that doesn't require padding
+        kwargs["config"].vision_config.pad_image_border = False
+
+        self.inner = LlavaForConditionalGeneration(**kwargs)
+        self.inner.multi_modal_projector = self.MULTIMODAL_PROJECTOR_TYPE(
+            kwargs["config"]
+        )
+        self.inner.get_image_feature = self.get_image_feature
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        """Extract features from image inputs.
+
+        Args:
+            items: List of MultimodalDataItem objects containing image data
+                Note that an item can be either "image" or "multi-images"
+
+        Returns:
+            torch.Tensor: features from image inputs, concatenated
+        """
+        features = []
+        for item in items:
+            # in each item, we assume pixel_values is always batched
+            pixel_values, image_sizes = item.pixel_values, item.image_sizes
+            image_outputs = self.vision_tower(
+                pixel_values, image_sizes, output_hidden_states=True
+            )
+            selected_image_feature = image_outputs.hidden_states[
+                self.vision_feature_layer
+            ]
+
+            if self.vision_feature_select_strategy in ["default", "patch"]:
+                selected_image_feature = selected_image_feature[:, 1:]
+            elif self.vision_feature_select_strategy == "full":
+                selected_image_feature = selected_image_feature
+            else:
+                raise ValueError(
+                    f"Unexpected select feature: {self.vision_feature_select_strategy}"
+                )
+            features.append(
+                self.multi_modal_projector(
+                    selected_image_feature.squeeze(0), image_sizes
+                )
+            )
+        ret = torch.cat(features, dim=0)
+        return ret
+
+    def __getattr__(self, name):
+        return getattr(self.inner, name)
+
+    def __hasattr__(self, name):
+        return hasattr(self.inner, name)
+
+    def __call__(self, *args, **kwargs):
+        return self.inner(*args, **kwargs)
+
+
+EntryClass = [MistralForCausalLM, Mistral3ForConditionalGeneration]

sglang/srt/models/mixtral.py

@@ -16,13 +16,15 @@
 # https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/mixtral.py#L1
 """Inference-only Mixtral model."""
 
-from typing import Iterable, Optional, Tuple
+import logging
+from typing import Iterable, Optional, Tuple, Union
 
 import torch
 from torch import nn
 from transformers import MixtralConfig
 
 from sglang.srt.distributed import (
+    get_pp_group,
     get_tensor_model_parallel_world_size,
     tensor_model_parallel_all_reduce,
 )
@@ -38,14 +40,17 @@ from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.radix_attention import RadixAttention
 from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
 from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
 from sglang.srt.managers.schedule_batch import global_server_args_dict
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
-from sglang.srt.utils import add_prefix
+from sglang.srt.utils import add_prefix, make_layers
+
+logger = logging.getLogger(__name__)
 
 
 class MixtralMoE(nn.Module):
@@ -257,24 +262,32 @@ class MixtralModel(nn.Module):
         super().__init__()
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size
+        self.pp_group = get_pp_group()
 
-        self.embed_tokens = VocabParallelEmbedding(
-            config.vocab_size,
-            config.hidden_size,
-            prefix=add_prefix("embed_tokens", prefix),
-        )
-        self.layers = nn.ModuleList(
-            [
-                MixtralDecoderLayer(
-                    config,
-                    i,
-                    quant_config=quant_config,
-                    prefix=add_prefix(f"layers.{i}", prefix),
-                )
-                for i in range(config.num_hidden_layers)
-            ]
+        if self.pp_group.is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                prefix=add_prefix("embed_tokens", prefix),
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        self.layers, self.start_layer, self.end_layer = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: MixtralDecoderLayer(
+                config=config, quant_config=quant_config, layer_id=idx, prefix=prefix
+            ),
+            pp_rank=self.pp_group.rank_in_group,
+            pp_size=self.pp_group.world_size,
+            prefix="layers",
+            return_tuple=True,
         )
-        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        if self.pp_group.is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer(return_tuple=True)
 
     def forward(
         self,
@@ -282,18 +295,35 @@ class MixtralModel(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
-    ) -> torch.Tensor:
-        if input_embeds is None:
-            hidden_states = self.embed_tokens(input_ids)
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+    ) -> Union[torch.Tensor, PPProxyTensors]:
+        if self.pp_group.is_first_rank:
+            if input_embeds is None:
+                hidden_states = self.embed_tokens(input_ids)
+            else:
+                hidden_states = input_embeds
+            residual = None
         else:
-            hidden_states = input_embeds
-        residual = None
-        for i in range(len(self.layers)):
+            assert pp_proxy_tensors is not None
+            hidden_states = pp_proxy_tensors["hidden_states"]
+            residual = pp_proxy_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
             layer = self.layers[i]
             hidden_states, residual = layer(
                 positions, hidden_states, forward_batch, residual
             )
-        hidden_states, _ = self.norm(hidden_states, residual)
+
+        if not self.pp_group.is_last_rank:
+            return PPProxyTensors(
+                {
+                    "hidden_states": hidden_states,
+                    "residual": residual,
+                }
+            )
+        else:
+            hidden_states, _ = self.norm(hidden_states, residual)
+
         return hidden_states
 
 
@@ -306,6 +336,7 @@ class MixtralForCausalLM(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
+        self.pp_group = get_pp_group()
         self.config = config
         self.quant_config = quant_config
         self.model = MixtralModel(
@@ -322,12 +353,31 @@ class MixtralForCausalLM(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
         input_embeds: torch.Tensor = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
     ) -> torch.Tensor:
-        hidden_states = self.model(input_ids, positions, forward_batch, input_embeds)
-        return self.logits_processor(
-            input_ids, hidden_states, self.lm_head, forward_batch
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            forward_batch,
+            input_embeds,
+            pp_proxy_tensors=pp_proxy_tensors,
         )
 
+        if self.pp_group.is_last_rank:
+            return self.logits_processor(
+                input_ids, hidden_states, self.lm_head, forward_batch
+            )
+        else:
+            return hidden_states
+
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
@@ -348,6 +398,17 @@ class MixtralForCausalLM(nn.Module):
 
         params_dict = dict(self.named_parameters())
         for name, loaded_weight in weights:
+            layer_id = get_layer_id(name)
+            if (
+                layer_id is not None
+                and hasattr(self.model, "start_layer")
+                and (
+                    layer_id < self.model.start_layer
+                    or layer_id >= self.model.end_layer
+                )
+            ):
+                continue
+
             if "rotary_emb.inv_freq" in name:
                 continue
 
@@ -398,11 +459,14 @@ class MixtralForCausalLM(nn.Module):
                     if name is None:
                         continue
 
-                    param = params_dict[name]
-                    weight_loader = getattr(
-                        param, "weight_loader", default_weight_loader
-                    )
-                    weight_loader(param, loaded_weight)
+                    if name in params_dict.keys():
+                        param = params_dict[name]
+                        weight_loader = getattr(
+                            param, "weight_loader", default_weight_loader
+                        )
+                        weight_loader(param, loaded_weight)
+                    else:
+                        logger.warning(f"Parameter {name} not found in params_dict")
 
 
 EntryClass = MixtralForCausalLM

sglang/srt/models/mllama.py

@@ -836,7 +836,6 @@ class MllamaForConditionalGeneration(nn.Module):
             prefix="multi_modal_projector",
         )
         self.logits_processor = LogitsProcessor(config.text_config)
-        self.capture_mode = False
 
     def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
         pixel_values = torch.cat(
@@ -865,7 +864,6 @@ class MllamaForConditionalGeneration(nn.Module):
                 pixel_values = torch.cat(
                     [item.pixel_values for item in mm_input.mm_items], dim=0
                 )
-                # max_num_images = max(max_num_images, sum(1 if item.is_image() else 0 for item in mm_input.items))
                 max_num_images = max(max_num_images, pixel_values.shape[1])
 
                 max_num_tiles = max(max_num_tiles, pixel_values.shape[2])
@@ -970,6 +968,8 @@ class MllamaForConditionalGeneration(nn.Module):
         positions: torch.Tensor,
         forward_batch: ForwardBatch,
     ) -> Union[Tuple, CausalLMOutputWithPast]:
+        from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
+
         batched_images, batched_ar_ids, batched_ar_mask, encoder_lens_need = (
             self._batch_image_inputs(forward_batch)
         )
@@ -978,7 +978,7 @@ class MllamaForConditionalGeneration(nn.Module):
         cross_attention_mask = None
         cross_attention_states = None
 
-        if self.capture_mode:
+        if get_is_capture_mode():
             # NOTE: when doing cuda graph capture, we do not want to skip cross attention
             # Make is a constant value to avoid cuda graph capture issue
             skip_cross_attention = False